The first supercomputers based on Nvidia’s H100 compute GPUs are yet to set records in terms of absolute performance, but they already show their might in terms of performance-per-watt.
Flatiron Institute’s Henri (opens in new tab) supercomputer, based on Intel’s Xeon Platinum 8362 (Ice Lake) and accelerated by Nvidia’s H100 compute GPUs this week, debuted in the Top500 and Green500 (opens in new tab) lists. In addition, it dethroned the AMD-powered Frontier Test and Development System, running AMD’s EPYC and Instinct MI250X hardware from the top of the Green500 list.
Lenovo built the Henri supercomputer, and it is currently the No. 405 most powerful system in the Top500 list with a Rmax performance of 2.04 FP64 PFLOPS, which is hardly impressive by itself. What is remarkable is that the machine consumes only 31 kW of power, demonstrating an energy efficiency of 65.091 GFLOPS/Watt, the world’s record. To put the number into context, the Frontier TDS machine hits 62.684 PFLOPS/W, the Frontier — the world’s fastest supercomputer — scores 52.227 PFLOPS/W, whereas the Lumi system achieves 58.021 PFLOPS/W.
The Henri machine is a relatively simplistic supercomputer by today’s standards: it uses Lenovo’s off-the-shelf air-cooled ThinkSystem SR670 V2 servers featuring Intel’s 32-core Xeon Platinum 8362 processors (5,920 cores in total) and 80 Nvidia’s H100 80GB PCIe cards based on the Hopper architecture. Of course, using air cooling for a relatively small system might have some other impact on its performance-per-watt results. However, Nvidia’s latest compute GPUs offer impressive performance in general.
“This supercomputer opens up opportunities for doing new kinds of science,” said Ian Fisk, co-director of the Flatiron Institute’s Scientific Computing Core. “This is a workhorse machine, and we’re going to let our researchers try new things and drive discoveries. […] [It offers] very high performance and very efficient without being particularly exotic. It only took a couple of people to load the system in. This kind of efficiency is now accessible to a lot more groups rather than just the largest supercomputing centers.”
Truth to be told, Nvidia-based supercomputers (in many cases comprising of standard servers) have been performance-per-watt champions in the Green500 list for some time, so it is logical to expect H100 to continue Nvidia’s winning spree here.
Meanwhile, AMD EPYC and Instinct MI250X-based machines are not outsiders in terms of performance-per-watt metrics, especially if you consider the scale of Frontier, Lumi, AdAstra, Setonix, and Dardel machines powered by AMD’s technology. In addition, six out of the top 10 supercomputers in the Green500 list utilize AMD’s CPUs and GPUs, three are accelerated by Nvidia’s compute GPUs, and one uses Intel’s Xeon Platinum 8260M-based nodes.
From what I'm reading, H100 PCIe cards have fp64 performance of 24 TFLOPS for 350 W (not sure if that's base or boost). By comparison AMD's MI210 PCIe card provides 13.3 or 22.6 fp64 TFLOPS (base vs. boost) at 300 W. If Nvidia's figures are base, then I'd say it checks out. If they're boost, then it sure looks close.
Anyway, that's a good comeback from Nvidia, especially when you consider they still have a 2:1 fp32:fp64 ratio. That said, their H100 is made on TSMC N4, whereas AMD's MI200-series uses N6. So, the tables will likely turn once again, when AMD counters with the MI300-series.